Auto-review applies to Shell, MCP, and Fetch calls. Allowlisted actions run immediately, sandboxable actions stay in the sandbox, and everything else is routed to a classifier subagent that can approve, redirect, or ask for human input.
This should cut approval fatigue for long-running agent sessions while keeping a review layer in place. It is a practical step toward more continuous agent work, especially for developers who are comfortable with policy but do not want constant interruptions.
Turn it on in Cursor Settings under Agents > Run Mode, then add custom instructions if you want to shape how the classifier behaves. Use it first on tasks with clear boundaries, like refactors, test runs, and repository-scoped maintenance.
Read Original Post →